Analyzing and Improving CPU and Energy Efficiency of PM File Systems

Persistent memory (PM) provides advantages such as direct access from the CPU at low latency, but it turned out that its performance is sensitive to certain access patterns. Particularly, the access latency increases when the PM is under parallel load. This leads not only to degraded PM performance, but slows down the entire system and wastes energy due to busy, but stalling CPUs. We propose an efficiency metric for measuring the CPU cost of a PM storage system. We evaluate our metric and the power consumption of existing file systems and explore PM-optimized memory copy routines to improve their efficiency. Finally, we propose an alternative system design where access to PM is mediated by an FPGA while allowing selective user space access to files.